Audit
During the public alpha, every prediction landed in a public, immutable ledger before kickoff — hashed, salted, and anchored to Bitcoin. That record is sealed and verifiable forever; the official SplitWinner ledger opens with its first anchor when the pipeline restarts.
Live
Live picks anchored
530
Conformal pass rate
13.4%
Anchored picks settled
70
Immutable since
May 20, 2026
Live · /v1/audit/summary + /v1/quant
Outside the sealed ledger — disclosed, never blended
The live pipeline has run since 2026-02-16. 926 settled picks sit outside the sealed ledger — most predate public anchoring, the rest fell on days without an anchor. Immutable database timestamps; verifiable to us, not verifiable to you — excluded from every figure above.
The record splits at the rebrand: the alpha's evidence is sealed and archived; the official chain carries the practice forward under a redesigned verifier. Every link on this page tells you which one it opens.
Sealed alpha · May 19 – Jun 25, 2026
29 anchors, both verifier generations, every disclosure stated in ALPHA.md — byte-identical to the record at sealing, Bitcoin attestations intact. This is where every existing receipt lives.
github.com/SplitWinner/audit_trail_alpha (opens in new tab)Official ledger
The redesigned protocol: one verifier for the whole chain forever, day-linked anchors, reports bound in, a public SPEC registry. Opens with its first MLB anchor; other sports join at their season starts.
github.com/SplitWinner/audit_trail (opens in new tab)§ 01Two records
We're new. We don't hide it. Here's the full chronology — characterisation, deployment, and the live record since. Both eras are public; we never blend the figures.
Before
The model run on historical games. Same pipeline that powers live today — but recorded after the games settled. Useful for characterising the model; never dressed up as a live call.
17,043 backtest predictions
The boundary
Audit trail deployed. From this day, every live morning pick on an anchored day was timestamped before kickoff, SHA-256 hashed, committed to a public GitHub ledger, and OpenTimestamps-anchored to Bitcoin.
May 20, 2026
After
Every alpha-era live pick on an anchored day is in the sealed public ledger — timestamped before kickoff, upgraded to a Bitcoin attestation, and refusable-by-design if a single byte moved. Picks outside the ledger are disclosed above, never blended in.
530 live picks anchored
§ 02What a tout can't do
Ten properties most prediction services skip. Each one of ours maps to a mechanism anyone can check.
| Property | Industry norm | SplitWinner |
|---|---|---|
| Predictions can be edited after the fact | Yes — usually a mutable table | No — append-only with DB trigger |
| Third-party-verifiable timestamp | No | Yes — -anchored daily |
| Independently auditable | No | Yes — public GitHub mirror |
| Model artifacts have a content hash on file | No | Yes — SHA-256 in immutable table |
| Training environment recorded | No | Yes — Python / library / CPU / BLAS captured |
| Inference byte-reproducible across runs | No | Yes — tier-3 determinism |
| BLAS thread pools pinned for determinism | No | Yes — set before Python starts |
| End-to-end reproducibility CI-verified | No | Planned — input capture shipped, harness pending |
| Backtest reads point-in-time data | Almost never | Yes — point-in-time, no lookahead leakage |
| Abstention is mechanically grounded | No — arbitrary confidence cutoffs | Yes — every prediction is ; what clears the bar is delivered, the rest is a disclosed skip |
§ 03How it works
From inference to public commitment, the whole flow is reproducible — and the verifier needs nothing from us.
01
The model emits a calibrated probability plus its conformal set.
02
The full prediction payload is hashed with SHA-256 to a stable fingerprint.
03
The hash is committed to the public audit_trail ledger before the game starts.
04
The full prediction is published to the same ledger before the game starts; the salted manifest hash binds it.
05
Anyone can re-hash the published prediction and check it matches the pre-game salted manifest.
§ 04How the rigor is enforced
What the build actually does — named, sourced, and verifiable. Each card is a mechanism that only exists when the rest is also real.
01
The audit hash is computed over a named-field subset of each prediction, canonicalised with json.dumps(sort_keys=True, separators=(",",":")), then SHA-256. Same bytes in, same hash out — across hosts, across years. The hash isn't asserted; it's re-derivable from the published row.
audit_trail row · field set documented under contract
02
Every training artifact lives in Supabase Storage at a path keyed by its own SHA-256 (models/[artifact_sha256].tar.gz). On every inference, the artifact is re-downloaded, re-hashed, and compared. Mismatch fails loudly — not a determinism bug; a bucket corruption or a swap.
audit_models.artifact_sha256 · audit_models.artifact_storage_path
03
Every audit_models row records python_version, package_versions (jsonb), blas_variant, and thread_env (jsonb). NULL on any of these is itself disclosed — those rows aren't valid reproducibility sources, and the daily CI refuses to use them as goldens.
audit_models columns · python_version · package_versions · blas_variant · thread_env
04
Every push and PR to the public audit_trail repo runs the integrity workflow. It blocks any modification, rename, or deletion of an anchored audit_trail row, model registration, or .ots proof — the append-only guarantee is a passing build, not a written promise. The workflow runs on every commit, not on a schedule.
audit_trail/.github/workflows/integrity.yml
05
Backtests don't read today's state of mutated tables. Three layers of mutation are explicitly audited — ingest UPSERTs, Redis in-place rewrites, in-path code reads — and seven specific code sites are named for as_of_date retrofit. Snapshots content-addressed by SHA-256.
Snapshot tables · mutation-site audit under contract
06
If post-calibration metrics on held-out data are worse than the raw model, calibration is rejected entirely. Stacking is leakage-aware — base models, meta-learner, and calibration each train on strictly separated folds. The rule is named, not a marketing line.
Calibration + stacking discipline · documented under contract
All six are mechanisms, not assertions. None of them work without the rest of the build also being real.
§ 05Activity
Each day the prediction pipeline ran, a manifest landed in the public ledger and was upgraded to a Bitcoin attestation. The sealed alpha holds 29 anchors (May 20 – Jun 25, 2026); the official ledger continues the practice from its first anchor.
Live anchors
530
Anchoring since
May 2026
Promotion rate
13.4%
Recent anchors
Live · syncing with /v1/audit/summary
95a29c4ccbaded85
Disclosed skip
Jun 25 · 12:00 UTC
52912fe37470f35e
Disclosed skip
Jun 25 · 12:00 UTC
656e8ef4dadb30e4
Disclosed skip
Jun 25 · 12:00 UTC
e0277ba8f6fafa15
Disclosed skip
Jun 25 · 12:00 UTC
a1fd43054638260f
Disclosed skip
Jun 25 · 12:00 UTC
9c0e05b33b8054fd
Disclosed skip
Jun 25 · 12:00 UTC
df8fc2d8c5e363f4
Disclosed skip
Jun 25 · 12:00 UTC
ca2059ce10e79bb1
Disclosed skip
Jun 25 · 12:00 UTC
5edc313fb630c54f
Disclosed skip
Jun 25 · 12:00 UTC
cfec7d82f712bd8a
Promoted
Jun 25 · 12:00 UTC
§ 07Verify it yourself
The verifier is plain Python. Clone, run, read the output.
git clone https://github.com/SplitWinner/audit_trail_alphacd audit_trail && python verify.py content --predictions sample/predictions_full.jsonAnchor timestamp · content hash · match status (per prediction)§ 08Uncertainty
Where the model commits to one outcome and where it hedges between two. Live and backtest side by side — the same procedure across both records.
Promotion rate (live)
13.4%
Live
Backtest
Math, not a hunch.
§ 09MonitoringLive only
Daily live calibration tracked over time. Vertical markers are retrains — drift triggers retraining, not the other way around. Backtest has no drift to track (it's a static replay).
§ 10Training cadence
Every model registration writes a row to the immutable audit_models ledger. The last few here — when each landed, what training-window it covers, and the commit SHA you can grep for in audit_trail.
Last registration
Jun 11, 2026
In last 30 days
0
Reproducible
100%
Recent registrations
Newest first. Source: the audit_models ledger written by the box pipeline.
mlbSP 2Recovered
Data through Dec 2026b6c98039
Jun 11, 2026
mlbSP 1Recovered
Data through Dec 2026b6c98039
Jun 11, 2026
nhlSP 2Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nhlSP 1Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nhlML 2Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nbaML 2Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nhlML 1Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nbaML 1Recovered
Data through Dec 2025b6c98039
Jun 11, 2026
nhlSP 2Recovered
Data through Dec 20258ca5360d
May 19, 2026
nhlSP 1Recovered
Data through Dec 20258ca5360d
May 19, 2026
§ 11We hide nothing
Honesty is the product. A buyer who finds a limitation we already disclosed trusts us more, not less.
Six axes, honest. The two BASIC ones are tracked, ordered, and owned in the public roadmap — not dressed up.
Integrity / verifiability
StrongImmutable ledger, Bitcoin anchoring, public verifier.
Operations
StrongOPERATIONS.md, single-operator reality disclosed.
Security
StrongKey custody, API hardening, read-only seam.
Quant rigor
StrongCalibration, conformal, point-in-time, determinism.
Data compliance
BasicData-lineage anchoring + formal docs not yet built.
Commercial
BasicNDA / MSA / SLA / pricing sheet / DPA not yet productized.
The conformal coverage target, per (sport, season_type). Every prediction row in the public audit_trail commits the applied value into its content_hash; every daily anchor publishes the policy block. If we change the rule, you'll see it on the next anchor.
01
The system is run by one operator today. Key custody, hosting, backups, recovery times, and known gaps are documented in audit_trail/OPERATIONS.md. No pretense of a SOC-2'd team we don't have.
02
The 16 production models predate the audit ledger. They were registered after-the-fact and are explicitly marked recovered=true on every row. The first anchor (May 2026) marks the start of unimpeachable live capture.
03
Two disclosed incidents in audit_trail/incidents/, under a 7-day disclosure protocol. The immutability guard prevents recurrence of the first; ingestion monitoring is the work for the second.
04
Data-lineage anchoring and formal data-handling docs are not yet built. NDA / MSA / SLA / pricing sheet / DPA are not yet productized. Tracked in the public roadmap; we don't claim what we haven't built.
§ 12Questions
What a vendor questionnaire asks, plainly answered.
A database we control can be edited, even if we promise not to. Bitcoin's block history isn't under our control — anyone can independently verify a hash was committed by a specific block. Same property regulators look for in tamper-evident logs.
The salt is held server-side (the `salt` column on `audit_anchors`) and revealed alongside the day's predictions once the anchor is published. Lose it before reveal and that day's manifest can't be re-verified; lose it after and verification still works from the published salt. The public repo carries the manifest hash and the OpenTimestamps proof; the integrity claim travels through Bitcoin even if our DB is gone.
Yes. The cadence is one anchor per prediction run, which is once per day while the pipeline runs. The integrity model doesn't care about cadence — every anchor is a standalone tamper-evident commitment.
You shouldn't. The whole point of the audit-trail design is to make trust unnecessary. The single-operator constraint is disclosed; the integrity primitives don't depend on the operator being honest after the fact.
Not productized yet. We document everything we'd commit to in audit_trail/OPERATIONS.md so a security team can evaluate before we sign anything formal.
No accounts, no payment, no permission. Clone, run, check.
https://github.com/SplitWinner/audit_trail_alpha (opens in new tab)If we vanished tomorrow, the proof would still stand.
A key in minutes — the full API surface, mocked plausibly-live, no card, no NDA. When it holds up on your desk, the 2026 beta cohort is a direct conversation: a limited set of desks, a fixed seasonal price in writing.
No card · no NDA · fixed price in writing